@gemstack/ai-sdk

@gemstack/ai-sdk is a provider-agnostic agent runtime. Define an agent once, then swap between Anthropic, OpenAI, Google, Ollama, Groq, DeepSeek, xAI, Mistral, Bedrock, and others by changing one model string. The engine handles tool calling, streaming, middleware hooks, structured output, multi-modal attachments, sub-agents, conversation memory, and a test fake. Its only required runtime dependency is zod; provider SDKs are optional peers you install per provider.

import { AiRegistry, AnthropicProvider, agent } from '@gemstack/ai-sdk'

AiRegistry.register(new AnthropicProvider({ apiKey: process.env.ANTHROPIC_API_KEY! }))
AiRegistry.setDefault('anthropic/claude-sonnet-4-6')

const response = await agent('You are a helpful assistant.')
  .prompt('Summarize the transformer architecture in one sentence.')

console.log(response.text)

Models are always provider/model; a bare model name throws. See /guide/installation for setup and /guide/first-agent for a full walkthrough.

Subpath exports

The engine works in any fetch-capable JS runtime (Node, browser, Electron, React Native): the main entry has zero static node:* imports.

Subpath	What it provides
.	Core: Agent, agent(), AI, tools, streaming, providers, middleware, structured output, run stores, testing fake
./node	Node-only filesystem helpers for attachments
./observers	Observer registry (aiObservers) for run telemetry
./chat-mentions	@slug agent routing helpers (parseMentions, buildMentionRoutingRule)
./gateway	Custom LLM-gateway / proxy adapter helpers
./eval	Eval framework (evalSuite, metrics, reporters)
./computer-use	Computer-use tool + executor
./react	React client bindings (useAgentRun)

The MCP bridge moved out to its own package, /packages/ai-mcp.

What's in the box

• Running agents - the three agent shapes, multi-step loops, sub-agents, suspend/resume, queued prompts, and middleware.
• Tools - toolDefinition().server(), scoped (multi-capability) tools, client tools, approval gates, and parallel execution.
• Streaming - chunk iteration, Server-Sent Events, and the Vercel AI SDK adapter.
• Structured output - typed, schema-validated responses.
• Memory - conversation persistence and cross-thread user memory behind neutral contracts.
• RAG - hosted vector stores and fileSearch retrieval.
• Providers - the provider catalog and per-provider configuration.
• Testing - the AiFake programmable mock and the eval harness.

For higher-level orchestration built on this runtime, see /packages/ai-autopilot; for Model Context Protocol bridging, see /packages/ai-mcp.

Where this came from

@gemstack/ai-sdk is the engine spun out of Rudder's @rudderjs/ai (carried forward from the 1.17.x line, renamed and re-versioned under the GemStack umbrella). @rudderjs/ai now re-exports this engine and adds the Rudder-specific bindings on top (the framework service provider, ORM-backed stores, and the make:agent / ai:eval CLI). Those bindings are not part of this package: everything documented here is what @gemstack/ai-sdk itself exports, usable with no framework.